| With the rapid development of consumer electronics,semiconductor,optical manufacture and other industries,high demands on the accuracy,efficiency,cost and reliability are required for specular surface reconstruction and defect inspection solutions.Phase Measuring Deflectometry and Structured-Light Modulation Analysis Technique have become competitive measurement solutions in the field of 3D measurement of specular surfaces,due to their advantages of simple system structure,high measurement accuracy and sensitivity.However,the related principles also make the techniques characterized by the difficulty of balancing accuracy and efficiency.Multi-frame demodulation methods such as the phase-shift method have higher accuracy,but require multiple shots of the tested specular surface,thus limiting the measurement efficiency;single-frame methods such as the Fourier transform method have lower accuracy and cause problems such as loss of detail information.To achieve dynamic measurements,researchers have also proposed measurement techniques with orthogonal composite grating encoding,but spectral aliasing,color crosstalk and Fourier transform also limit the accuracy of the related methods.Inspired by the good performance of deep learning in recent years in the field of computational imaging including fringe analysis,this thesis presents detailed study of single-frame phase measurement,modulation measurement and dynamic 3D reconstruction methods based on deep learning for the pain points of Phase Measuring Deflectometry and Structured-Light Modulation Analysis Technique.In this thesis,the basic principles and key algorithms of Phase Measuring Deflectometry,as well as the development and advantages of deep learning techniques are described in detail.Combined with the neural network structures that have performed well in recent years,this thesis designs an improved U-NET for fringe analysis mission.For single-frame phase retrieval and modulation measurement,a distorted fringewrapped phase dataset and a distorted fringe-modulation dataset are produced by traditional phase-shift PMD techniques,and a model is obtained by training the network with the dataset.The model is used to complete single-frame phase retrieval and singleframe modulation measurement.For dynamic measurement of specular surfaces,a distorted orthogonal fringe-wrapped phase dataset is produced to train the proposed network.High-accuracy phase retrieval in both directions is successfully completed by the trained model,and single-frame 3D reconstruction and dynamic measurements of the tested specular surface are realized.The experimental results demonstrate that both the single-frame phase retrieval and modulation retrieval methods proposed in this thesis can obtain high accuracy results from a single-frame distorted fringe pattern,and their accuracy is close to that of the tenstep phase-shift method.The dynamic measuring method proposed in this thesis achieves phase prediction in both directions from a single-frame distorted orthogonal fringe,and the results are also superior to the 2D WFR method,reaching that of the ten-step phase shift method.The relative average error is about 1.17% for dynamic specular surfaces measured by the proposed method.The relevant experiments validate the excellent performance of the proposed method for high-speed and high-precision 3D measurement of specular surface. |
